Angus McDonald (@AngusRmcdonald) header
Angus McDonald (@AngusRmcdonald) profile picture

Building the AI Company Brain

Sydney, Australia
Joined April 2014
Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)Photo by Angus McDonald (@AngusRmcdonald)
Angus McDonald retweeted
Replying to @karpathy
13
74
8
1,789
93,318
We ran gpt-5.4-nano through 240 computation problems across 10 domains. Raw LLM: 37.9% correct Provider tools: 56.7% Euclid MCP tools: 74.2% Three domains hit 100%. Here's what we learned. (thread)
1
83
Results across all 10 domains. Sorted by where Euclid helps most. Regex is the interesting outlier: 100% raw, 58.3% with tools. LLMs are already pattern machines. Not every domain needs a tool.
1
57
1/13: LLMs are terrible at maths. Not sometimes. Reliably. Ask any frontier model to calculate NPV or convert a hex colour, it'll confidently give you the wrong answer. Fine for chat. Dealbreaker for agents. I'm building a fix. 🧵
1
15
12/13: Full benchmark results across multiple models coming soon. Follow if you're building agents and care about accuracy over vibes.
1
8